Professional Documents
Culture Documents
Approximating Network Centrality Measures Using Node Embedding and Machine Learning
Approximating Network Centrality Measures Using Node Embedding and Machine Learning
Abstract
Extracting information from real-world large networks is a key challenge nowadays. For instance,
computing a node centrality may become unfeasible depending on the intended centrality due to
its computational cost. One solution is to develop fast methods capable of approximating network
centralities. Here, we propose an approach for efficiently approximating node centralities for large
networks using Neural Networks and Graph Embedding techniques. Our proposed model, entitled
arXiv:2006.16392v4 [cs.SI] 1 Nov 2020
Network Centrality Approximation using Graph Embedding (NCA-GE), uses the adjacency matrix
of a graph and a set of features for each node (here, we use only the degree) as input and computes
the approximate desired centrality rank for every node. NCA-GE has a time complexity of O(|E|), E
being the set of edges of a graph, making it suitable for large networks. NCA-GE also trains pretty
fast, requiring only a set of a thousand small synthetic scale-free graphs (ranging from 100 to 1000
nodes each), and it works well for different node centralities, network sizes, and topologies. Finally,
we compare our approach to the state-of-the-art method that approximates centrality ranks using
the degree and eigenvector centralities as input, where we show that the NCA-GE outperforms the
former in a variety of scenarios.
1 Introduction
Networks are present in several real-world applications spread among different disciplines, such as biology,
mathematics, sociology, and computer science, just to name a few. Therefore, network analysis is a crucial
tool for extracting relevant information. However, this analysis may be hindered when performed over
large complex networks, given the high computational cost for some network analysis methods. This
problem is especially challenging when we consider that nowadays many networks extracted from real-
world applications are usually large in scale.
Among the most used approaches for network analysis, node centrality is an essential concept related
to the assessment of the relative importance of nodes given some criteria of importance. There are
several different centrality measures, each with their own definition of a central (i.e., important) node. For
example, the most basic node centrality is the degree centrality which measures the number of connections
a node has as a proxy for its relative importance in the network. In this case, a node is considered central
when it has many connections. Other examples of common centrality measures are betweenness, closeness,
and eigenvector centralities, among others. Centrality measures have already been successfully applied to
many real-world problems, as for example: (i) optimization of computer networks [1]; (ii) transportation
networks [2]; (iii) biological networks [3]; (iv) recommendation systems [4]; (v) reinforcement learning [5];
and (vi) social networks [6], to name a few domains of application. For more applications, refer to [7, 8]
for a more in-depth review of real-world examples of network centrality usage.
The size of many complex networks found in real-world applications is large. Depending on the size of
a network, the direct calculation of a given network centrality measure may become unfeasible due to the
required high computational cost. Therefore, many solutions have been proposed for approximating or
even calculating the actual centrality value for large networks in a reasonable time, given some constraint
or specific target. The initial solutions focused on sampling techniques used for static graphs [9–12] or
parallel algorithms for specific centrality measures [13]. More recently, researchers have focused on the
problem of approximating network centrality for evolving graphs [14]. Another recent approach is to use
a Neural Network (NN) to predict node centrality based on lower-order node features [7, 15, 16], which is
the approach that inspired our work.
∗ M. R. F. Mendonça, A. M. S. Barreto and A. Ziviani are with the National Laboratory for Scientific Computing (LNCC),
London, UK.
1
2 NODE CENTRALITY MEASURES
In this paper, we focus on the problem of approximating node centrality measures for large complex
networks using node embedding and Machine Learning. Node (graph) embedding [17, 18] comprises
methods that use neural networks in order to build an abstract representation of a node or a graph,
encoding higher-order representations in these low dimension embeddings (check Section 3 for more
details). We propose here the Network Centrality Approximation using Graph Embedding (NCA-GE)
model, a model capable of predicting the rank of the desired centrality for all nodes in a graph of any
dimension and topology. To achieve this, our model only requires the adjacency matrix and a set of
features for each node in order to predict its value for other centrality measures, whereas other node
features are inferred by the node embedding. The set of features reported in this work is composed
simply of the degree centrality of all nodes, but any other node feature is allowed. We train our model
over a set of artificial complex networks. The model can be thus used to predict the network centrality
value of a given node in larger networks.
The main contributions of the NCA-GE model with respect to other similar previous works [7, 15, 16]
are:
• The NCA-GE model constitutes a more general framework that works with any graph embed-
ding method, making it highly customizable since it can harness the advantages of different graph
embedding methods in order to adapt the model for different network topologies.
• The NCA-GE model, as investigated here, requires less node features since it uses only the node
degree as a feature, whereas the current state-of-the-art model [7] uses the degree and the eigenvector
centrality to this end.
• The NCA-GE model outperforms the current state-of-the-art model [7] by providing more accurate
centrality rankings for different network topologies.
The source code for the all the presented models is publicly available on GitHub.1
The remainder of this paper is organized as follows. In Section 2, we present the basic concepts
of node centrality, node embedding, and neural networks, which compose the foundations of this work.
We then present some of the related work on node embedding and network centrality approximation in
Section 3. In Section 4, we introduce our proposed NCA-GE model, followed by the description of the
experiments and obtained results in Section 5. Finally, we conclude the paper and discuss some future
work in Section 6.
2
2 NODE CENTRALITY MEASURES
nodes. The eigenvector centrality of a node is thus proportional to the sum of the eigenvector centralities
of its neighbor nodes. Therefore, central nodes are usually connected to other central nodes or are
connected to a relatively large number of low centrality nodes. The most central nodes typically have
both. This is useful to identify influential nodes in a network or nodes connected to other influential
nodes. The eigenvector centrality xi of node i is formally defined as:
1X
xi = aij xj , (1)
λ
j∈V
where λ is an eigenvalue of the adjacency matrix A associated with the eigenvector x. Therefore, the
eigenvector centrality of each node in a graph can be calculated by solving the eigenvector problem
Ax = λx. (2)
The two centralities presented so far are connectivity-based centralities, i.e., they use node connectivity
to measure its importance. Connectivity-based centralities, which are not limited to the two previously
presented centralities, are computationally less expensive to be calculated.
We now present some shortest-path based centralities, which use minimum paths between nodes in
order to measure their importance. Hence, these centralities are far more expensive computationally,
since they require computing the minimum distance between all pairs of nodes. Check Grando et al. [7]
for more details in the computational time complexity for different types of node centralities.
where δ(i, j) is the distance between nodes i and j. A high closeness centrality means that the node is
on average closer to all other nodes in the network, i.e., the node with high closeness centrality can on
average reach other nodes or be reached by these nodes in fewer steps.
where 1/δ(i, j) = 0 if there is no path between i and j. Note that this allows infinity distances (for
unconnected graphs), which is not allowed for closeness centrality.
where σst (i) is the number of shortest paths between nodes s and t that passes through i and σst is the
total number of shortest paths between s and t. Therefore, betweenness centrality measures the extent
to which a node lies on paths between other nodes. Nodes with high betweenness may have considerable
influence within a network due to their control over information passing between other nodes.
3
3 RELATED WORK
GCN
Relu
x
Reduce
Sum Nx1 NxF
NxN FxF
S2VEC
Relu
NxF x NxF
+ NxF
NxC FxF
Relu
CxF NxF
Graph
Embedding
(c) NCA-GE.
Figure 1: Architecture used for the graph embedding methods and the proposed model: (a) Graph
Convolutional Network (GCN), (b) Structure2Vec (S2VEC), and the Network Centrality Approximation
using Graph Embedding (NCA-GE) model. The inputs are represented in yellow, the node embedding
matrix is in light red, hidden layers are in blue, intermediate matrices are in green, and the output is
represented in red. N represents the number of nodes in the graph being processed in the batch, C is the
number of features used in the feature matrix, F is the number of features used in the embedding, and
1 1
B is the batch size. For each batch, we require the normalized adjacency matrix D̃− 2 ÃD̃− 2 H (l) W (l)
matrix (for the GCN) or the common adjacency matrix A (for the S2VEC), the feature matrix containing
the features of all nodes, and the node id vector containing the id of all nodes in the batch. The output is
a single value for each node in the batch, representing the normalized rank of the desired centrality. The
proposed model works with any graph embedding method, showcased by the generic “Graph Embedding”
box at the to left corner, which outputs the embedding matrix.
3 Related Work
In this section, we briefly review some of the recent advances in node centrality prediction and node
embedding, which are the techniques that lay the foundation for this work.
4
3.1 Node and Graph Embedding 3 RELATED WORK
where
P ˜ σ is an activation function, Ã = A + I is the adjacency matrix plus the identity matrix I, D˜ii =
l 0
j Aij , and W is the parameter matrix of layer l. Here, H is given by the input matrix FM and the
l
output matrix H corresponds to the matrix H of the last layer. This formulation is especially interesting
because it is isomorphic graph invariant, i.e., the result is the same regardless of the order of the nodes
(which yields different adjacency and diagonal matrices). Figure 1(a) shows the schematics of the GCN
model. For more implementation details about the GCN, check Section 4.1.
Structure2Vec [22] is a recent graph embedding approach where each node is codified by considering
a given information of the node and the embedding of all of its neighborhood. It is a recursive approach
that updates each node embedding through a given number of layers. After L layers, the node embedding
hL
i of node i in layer L considers the topological information of all nodes in its L-hop neighborhood. If
we run the method with dm layers, where dm is the diameter of the graph, then the node embedding
considers the topological information of the entire graph. Since the diameter of real-world networks is
usually small compared with the network size, then the number of layers required to embed the whole
graph is also small. Formally, the embedding hi of a node i at layer l + 1 is defined as a non-linear
function f that receives three parameters:
5
3.2 Network Centrality Prediction 4 NCA-GE MODEL
X
hl+1
i = f xi , hln , w(i, n)n∈N (i) , (7)
n∈N (i)
where f is any nonlinear function (such as a neural network for example), N (i) is the set of neighbors
of node i, xi is any relevant node information that should be embedded, and w(i, n) is the weight of the
edge connecting nodes i and n (in the unweighted setup, such as the one used here, w(i, n) = 1 for all
existing edges and w(i, n) = 0 for non-existing edges). The node embeddings are then summed in order
to generate a graph embedding, which is used for graph classification problems. h0i is initialized with
zeros for all nodes i. Figure 1(b) shows the Structure2Vec model. More details about the implementation
of Structure2Vec are discussed in Section 4.2.
Hanjun et al. [23] adopt a similar approach, also based on the Structure2Vec embedding, but instead
of using it for classification problems, the authors adapt the method to work with reinforcement learning
in order to solve combinatorial graph problems.
Here, we use the GCN and Structure2Vec methods for our experiments.
4 NCA-GE Model
We present here the Network Centrality Approximation using Graph Embedding (NCA-GE) model, capa-
ble of approximating any node centrality using only the degree centrality as feature input. The approach
outlined in this paper is inspired by two distinct approaches: (i) the NN node centrality prediction model
recently introduced by Grando et al. [7]; and (ii) modern graph embedding techniques, such as the Graph
Convolutional Network (GCN) [21] and Structure2Vec [22]. The proposed idea is to codify each node
into an embedding vector calculated using only the degree centrality as a feature of a given node. We
6
4.1 GCN: Implementation Details 4 NCA-GE MODEL
then pass these embeddings to a NN model (Figure 1(c)), which outputs an approximation of the desired
centrality. The output centrality can be any centrality other than the degree (since it is already used as
input). The NCA-GE can use any graph embedding technique, but here we explore using two specific
methods: GCN and Structure2Vec. Our contributions are three-fold: (i) our approach achieves better
results for artificial and real-world networks when compared with previous work [7], (ii) the NCA-GE
uses less information as input, making it easier to be executed, as we show in Section 5.3, and (iii) our
model is more customizable in the sense that it works with any graph embedding method, allowing it
to be optimized for different scenarios (different network topologies, for example) using different graph
embedding methods.
In the remainder of this section, we discuss our method in more detail. We present the implementation
details for the GCN and Structure2Vec embedding methods used. The training process is also discussed.
We also present the implementation details of the baseline model used to compare with our approach.
The training parameters used for each of the models presented here are shown in Table 1.
Table 1: Parameters used for training the proposed model and the baseline model. GE Layers refers
to the number of iterations are used to compute the embedding matrix (Figure 1(a) and 1(b)); η is the
learning rate used; β is the learning rate decay; λ refers to the weight given to the regularization term
in the loss function; C is the number of input features used; and F is the size of the embedding vectors
generated.
GE Layers η β min η λ C F Batch Size
NCA-GE (GCN) 2 0.001 0.999 0.0001 0.01 1 128 128
NCA-GE (S2VEC) 2 0.001 0.999 0.0001 0.1 1 128 128
Baseline - 0.01 0.999 0.0001 0.001 2 128 128
where relu(.) is the rectified non-linearity function and the remaining variables are the same as in Eq. 6.
We used 2 layers, where W 1 ∈ IRCxF , W 2 ∈ IRFxF , C being the number of initial features and F being
the embedding size. As already mentioned, here we use only a single feature: the normalized degree rank
of a node, thus, C = 1.
N
X
di = Aij , ∀i = 1, ..., N (9)
j=1
X1 = dW 1 (10)
3
X3 = FM W (11)
0
H = relu(X1 + X3 ) (12)
H l+1 = relu AH l W 2 + X1 + X3 ,
(13)
7
4.3 NCA-GE 4 NCA-GE MODEL
where d ∈ IRNx1 is the degree vector, W 1 , W 2 ∈ IRFxF and W 3 ∈ IRCxF are the weight matrices, and
H l is the embedding matrix at layer l. We used 2 layers, therefore, the final embedding matrix is given
by H 2 . Notice that Structure2Vec uses the plain adjacency matrix A, unlike the GCN, which uses a
normalized version of A with self-loops.
4.3 NCA-GE
After computing the embedding matrix H (using whichever graph embedding method), we select only
the embeddings of the nodes being analyzed in the current batch. The selected nodes are given by:
HD = ID × H, (14)
where Id ∈ IRBxN is the one-hot encoding of the id of the nodes in the batch and B represents the size of
the batch. The centrality prediction outputted by the NN is then given by:
Y = relu(relu(relu(HD W F 1 )W F 2 )W F 3 )W F 4 , (15)
F1 F2 F3 F4 FxF
where W , W , W , W ∈ IR are the weight matrices.
As we can see from Figure 1, all weight matrices used are independent of the graph size N . This
allows our approach to be used to predict node centralities for graphs of any size. We can also notice
that both graph embedding models used here (GCN and Structure2Vec) require the complete adjacency
matrix A and the one-hot encoding of the nodes in the current batch. To achieve a fast training time
with our model, we use sparse tensor representations implemented in Tensorflow.
The feature input of the NCA-GE is the normalized rank value of the degree centrality of a given node
and the output is the normalized value of the rank of the desired higher-order centrality, both within
the range [0, 1]. Therefore, the only feature information required by our model is the degree centrality
of each node, which is a very low cost and simple centrality to be obtained. We experienced using the
normalized value of the exact degree centrality as an input, but the results were not as good as using the
normalized value of the degree rank, even when we intended to predict the normalized value of the exact
centrality as output.
The error function used for our NN model is the mean squared error (MSE) between the predicted
centrality and the expected value. We also used an L2 regularization function to avoid over-fitting.
Therefore, the adopted loss function is:
LT OT AL = LM SE + λLREGU L , (16)
where LM SE is the loss function for the Mean Squared Error, LREGU L is the L2 regularization loss, and
λ is a variable that controls the weight given to the regularization term. Regularization loss helps to
avoid the model to become too overfitted to the training data, losing its generalization capabilities. We
used the Adam algorithm as the optimization function, the gradients were truncated in the range [-1, 1],
and the batch size was set to 128.
η = ηβ, (17)
where η is the learning rate and β refers to the learning rate decay coefficient. The learning rate is
decayed until it reaches a minimum value. Check Table 1 for a full list of the parameters used. Finally,
the gradients used to update the weight matrices are clipped in the range of [-1, 1] in order to avoid
instabilities during training.
8
4.5 Baseline Model: Implementation Details 5 PERFORMANCE EVALUATION
5 Performance Evaluation
As detailed in Section 4.4, we train the proposed NCA-GE model by iterating over all nodes in a set
of training graphs. Each node fed to the model is considered a step. The results presented here in this
section for the NCA-GE model were all generated after training for 4 million steps, while the baseline
model was trained for 1 million steps.
As stated in Grando et al. [7], the MSE is not a reliable metric to evaluate how well a model that
predicts the centrality rank of a node is performing. This is due to the fact that we are interested
in a model that predicts the correct rank ordering of a given centrality measure for all nodes, not the
resulting absolute rank value. Therefore, we follow the methodology adopted by Grando et al. [7] and use
the Kendall τ -b [26] rank correlation coefficient to determine the quality of the NCA-GE rank prediction
for network centralities. The Kendall τ -b coefficient varies in the range of [-1, 1] and it measures the
correlation between two ranking lists. A value of -1 indicates that the two lists are fully reversed (that is,
if index X is ranked first in one list, it is ranked last in the other, and so on for the other rank positions).
A value of 0 means that there is no correlation, whereas a value of 1 indicates that both lists fully agree
with the ranks of every node. Therefore, to measure the performance of a prediction model, we first build
two rank lists: (i) a list in which the value in index i indicates the true normalized rank of node i, and
(ii) a similar list that stores the predicted normalized rank. We then measure the Kendall τ -b coefficient
between these two lists, where a value near 1 indicates good performance since the prediction model
managed to rank most of the nodes correctly.
We tested the NCA-GE model with four different node centralities: (i) betweenness, (ii) closeness,
(iii) harmonic, and (iv) eigenvector. The first three centralities are very computationally expensive, as
discussed in Section 2, while the fourth centrality is relatively cheap, but we experimented with it in
order to show the effectiveness of the NCA-GE model.
In this section, we detail several experiments that we performed to showcase the efficiency of the
proposed NCA-GE approach using the GCN or Structure2Vec graph embedding methods. We use as
a baseline to compare our results the model recently proposed by Grando et al. [7] (see Section 4.5).
We evaluate and compare our approach in a series of artificial complex networks generated by different
generation models and a set of real-world networks. We also analyze the time complexity of both baseline
and NCA-GE models. All experiments shown here for the NCA-GE model were generated using only
the normalized rank degree centrality as feature input for each node (as previously stated in Section 4),
while the results for the baseline model were generated using the normalized rank degree and eigenvector
centralities as input. These experiments aim to show that the NCA-GE achieves comparable or better
results when compared with the baseline model, and it manages this by using only the degree centrality
as a feature input.
9
5.1 Experimental Results with Synthetic Networks 5 PERFORMANCE EVALUATION
Figure 2: Mean Kendall coefficient for the synthetic network for the baseline, NCA-GE (S2VEC), and
NCA-GE (GCN) models. Each result is the average of 5 executions. We separated the results based on 4
different test sets: small-world set (SW), scale-free set (SF), random graphs set (RND), and a mixed set
containing all previous 3 graph types (MIX). Note that we do not compare the results for the Eigenvector
centrality with the baseline model since it can not predict this centrality, given that this centrality is one
of its inputs. The standard deviation was not included in the graph as it presented very small values for
all three models presented (values ranging from 10−3 to 10−5 ).
10
5.1 Experimental Results with Synthetic Networks 5 PERFORMANCE EVALUATION
Figure 3: Comparison of the mean Kendall Rank Coefficient for the real-world networks described in
Table 2 when using the baseline, NCA-GE (S2VEC), and NCA-GE (GCN). These results are computed
using an average of 5 executions. Note that we do not compare the results for the Eigenvector centrality
with the baseline model since it can not predict this centrality, given that this centrality is one of its
inputs. The standard deviation was not included in the graph as it presented very small values for all
three models presented (values ranging from 10−3 to 10−5 ).
cial scale-free networks, which makes sense since the training set used for all experiments was comprised
entirely of artificial scale-free networks. We can also see that the NCA-GE (S2VEC) performs as well
as the baseline model for the betweenness centrality while presenting a better result for the closeness
and harmonic centralities when compared to the baseline for all types of network topologies. This shows
just how powerful the NCA-GE with Structure2Vec is: It manages to achieve equal to better results
when compared with the baseline model without using any higher-order feature, such as the eigenvector
centrality. Regarding the NCA-GE using the GCN embedding, we can note that it did not excel as its
Structure2Vec counterpart, although it still managed to achieve comparable results to the baseline model
for the closeness and harmonic centralities, and almost equal results for betweenness. The results for the
eigenvector centrality evidences the superiority of the NCA-GE using Structure2Vec over the GCN for
computing the node embeddings.
As a final remark, in Figure 2, we can observe that the results for closeness and harmonic centralities
are very similar. This phenomenon can be explained by the similarities in how closeness and harmonic
centralities are computed. We can also note that the lowest obtained Kendall coefficients are for the
betweenness centrality. This is likely due to the fact that closeness and harmonic centralities are more
stable across different topologies, since this centrality tends to have low values and low deviation. On the
other hand, the distribution of values for the betweenness centrality tends to variate more depending on
the graph topology. This can be noted by the high Kendall coefficients for all models for the SF topology,
which was also used during training, but low values for the remaining topologies.
11
5.2 Experimental Results with Real-World Networks 5 PERFORMANCE EVALUATION
Figure 4: Comparison of the execution time for the real graphs using the baseline and NCA-GE models
in relation to the number of edges in the network. The Y-axis represents the time required to process
the centrality for all nodes in a graph. The X-axis represents the number of edges in the network.
These values represent the average of 5 executions. The error bars for each point represent the standard
deviation.
Figure 3 presents the results of the average Kendall τ -b coefficient of all four centralities for the
real networks (Table 2) using the baseline, the NCA-GE (GCN), and NCA-GE (S2VEC) models. The
obtained results for the real networks are somewhat similar to the obtained results for the synthetic test
sets: The NCA-GE using Structure2Vec managed to achieve equal to better results for all networks and
all centralities (except eigenvector centrality) when compared with the baseline model. Once again, these
results were obtained using only the normalized degree rank as input feature, contrasting to the baseline
model, which uses the degree and eigenvector centralities as input features. The results for NCA-GE
using GCN were also similar to the ones obtained for the synthetic test sets, where it achieved similar
results to the baseline model for the betweenness centrality, and slightly worse to equal results for the
closeness and harmonic centralities. For the eigenvector centrality, we can see a better performance of
the NCA-GE with GCN when compared with the obtained results for the test sets: Here it manages to
achieve slightly better results than the NCA-GE with Structure2Vec for some graphs, although the latter
still outperforms the former in the majority of the real-world graphs.
12
5.3 Time Complexity Analysis 5 PERFORMANCE EVALUATION
Table 3: Mean execution time (s) and standard deviation for processing the betweenness centrality using
the baseline and NCA-GE models for each of the real networks used for validation. These values represent
the average of 5 executions.
Graph Baseline NCA-GE
Email 0.024 ± 0.001 0.018 ± 0.001
FB 0.164 ± 0.007 0.141 ± 0.002
RO 0.420 ± 0.022 0.315 ± 0.033
HR 1.621 ± 0.074 1.200 ± 0.067
HU 0.701 ± 0.016 0.543 ± 0.008
GN 30 0.326 ± 0.011 0.238 ± 0.029
GN 31 0.615 ± 0.060 0.449 ± 0.006
SDot 1.707 ± 0.042 1.366 ± 0.063
The NCA-GE model requires the degree centrality and the sparse representation of the adjacency
matrix as input, which has a time complexity of O(N ) and O(|E|), respectively. After that, the model
performs a series of matrix multiplications, where these operations are bounded by the matrix operations
using the adjacency matrix A, with dimensions N ×N , resulting in a time complexity of O(N 2 ). However,
since we use sparse matrix representations, this complexity becomes associated with the density of the
matrix, i.e, O(|E|). Note that the NCA-GE computes the target centrality of all nodes in a single pass.
Therefore, the overall complexity of the NCA-GE model is given by O(N + |E| + cp |E|) = O(|E|), where
cp is a constant reflecting the number of operations performed by the proposed NCA-GE model.
Figure 4 presents the time required to process the centrality rank of every node in relation to the
number of edges for the real-world networks using the baseline and NCA-GE with Structure2Vec. Table 3
presents the execution time for each of the networks we use. As we can see, the time required to process a
network of both models grows linearly with the number of edges, confirming our time complexity analysis
of O(|E|). The processing time gap between the baseline model and the NCA-GE can be explained by the
former’s time complexity (O(k|E|)) being greater than the latter (O(|E|)). Since this difference is given
only by the number of iterations k of the power method, this value still maintains the behavior of O(|E|).
This makes the processing time of the baseline model to follow the same behavior of the NCA-GE, albeit
with a slight shift in the time axis, as depicted in Figure 4.
13
REFERENCES
Acknowledgment
This work has been partially supported by CAPES, CNPq, FAPERJ, and FAPESP. Authors also ac-
knowledge the INCT in Data Science – INCT-CiD.
References
[1] L. Maccari and R. L. Cigno, “Pop-routing: Centrality-based tuning of control messages for faster
route convergence,” in IEEE INFOCOM 2016 - The 35th Annual IEEE International Conference
on Computer Communications, 2016, pp. 1–9.
[2] K. Wehmuth, B. B. A. Costa, J. V. M. Bechara, and A. Ziviani, “A multilayer and time-varying
structural analysis of the brazilian air transportation network,” in Proceedings of the Latin America
Data Science Workshop, 2018, pp. 57–64.
[3] M. Rubinov and O. Sporns, “Complex network measures of brain connectivity: Uses and interpre-
tations,” NeuroImage, vol. 52, no. 3, pp. 1059 – 1069, 2010, computational Models of the Brain.
[4] R. Ying, R. He, K. Chen, P. Eksombatchai, W. L. Hamilton, and J. Leskovec, “Graph
convolutional neural networks for web-scale recommender systems,” in Proceedings of the
24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining,
ser. KDD ’18. New York, NY, USA: ACM, 2018, pp. 974–983. [Online]. Available:
http://doi.acm.org/10.1145/3219819.3219890
[5] M. R. F. Mendonça, A. Ziviani, and A. M. S. Barreto, “Graph-based skill acquisition for reinforce-
ment learning,” ACM Computing Surveys, vol. 52, no. 1, pp. 6:1–6:26, 2019.
14
REFERENCES REFERENCES
[6] J. A. Danowski and N. T. Cepela, “Automatic mapping of social networks of actors from text corpora:
Time series analysis,” in 2009 International Conference on Advances in Social Network Analysis and
Mining, 2009, pp. 137–142.
[7] F. Grando, L. Z. Granville, and L. C. Lamb, “Machine learning in network centrality measures:
Tutorial and outlook,” ACM Computing Surveys, vol. 51, no. 5, pp. 102:1–102:32, Oct. 2018.
[8] K. Das, S. Samanta, and M. Pal, “Study on centrality measures in social networks: a survey,” Social
Network Analysis and Mining, vol. 8, no. 1, p. 13, Feb 2018.
[18] P. Cui, X. Wang, J. Pei, and W. Zhu, “A survey on network embedding,” IEEE Transactions on
Knowledge & Data Engineering, 2019.
[19] P. Boldi and S. Vigna, “Axioms for centrality,” Internet Mathematics, vol. 10, no. 3-4, pp. 222–262,
2014. [Online]. Available: https://doi.org/10.1080/15427951.2013.865686
[20] D. Wang, P. Cui, and W. Zhu, “Structural deep network embedding,” in Proceedings of the 22Nd
ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, ser. KDD ’16.
ACM, 2016, pp. 1225–1234.
[21] T. N. Kipf and M. Welling, “Semi-supervised classification with graph convolutional networks,”
CoRR, vol. abs/1609.02907, 2016. [Online]. Available: http://arxiv.org/abs/1609.02907
[22] H. Dai, B. Dai, and L. Song, “Discriminative embeddings of latent variable models for structured
data,” in Proceedings of the 33rd International Conference on International Conference on Machine
Learning - Volume 48, ser. ICML’16. JMLR.org, 2016, pp. 2702–2711.
[23] E. Khalil, H. Dai, Y. Zhang, B. Dilkina, and L. Song, “Learning combinatorial optimization algo-
rithms over graphs,” in Advances in Neural Information Processing Systems 30 (NIPS). Curran
Associates, Inc., 2017, pp. 6348–6358.
15
REFERENCES REFERENCES
[27] P. Erdös and A. Rényi, “On random graphs i,” Publ. Math. Debrecen, vol. 6, p. 290, 1959.
[28] D. J. Watts and S. H. Strogatz, “Collective dynamics of ’small-world’ networks,” Nature, vol. 393,
no. 6684, pp. 440–442, Jun. 1998.
[29] J. Leskovec and A. Krevl, “SNAP Datasets: Stanford large network dataset collection,” http://snap.
stanford.edu/data, Jun. 2014.
[30] W. Richards and A. Seary, “Eigen analysis of networks,” Journal of Social Structure, vol. 1, no. 2,
pp. 1–17, 2000.
[31] E. C. Costa, A. B. Vieira, K. Wehmuth, A. Ziviani, and A. P. C. Da Silva, “Time centrality in
dynamic complex networks,” Advances in Complex Systems, vol. 18, no. 07n08, p. 1550023, 2015.
[32] P. Goyal, S. R. Chhetri, and A. Canedo, “dyngraph2vec: Capturing network dynamics using dynamic
graph representation learning,” Knowledge-Based Systems, 2019.
[33] K. Berahmand, A. Bouyer, and N. Samadi, “A new local and multidimensional ranking measure to
detect spreaders in social networks,” Computing, vol. 101, no. 11, pp. 1711–1733, 2019.
[34] C. Magnien and F. Tarissan, “Time evolution of the importance of nodes in dynamic networks,” in
2015 IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining
(ASONAM), 2015, pp. 1200–1207.
16